I think in multi-threaded code ref-counting usually uses atomic_fetch_add instructions. This results in a memory barrier and is much more expensive that you might think.